Poisoning Agentic Alpha attacks live LLM trading crews using only the data and prompts they already consume
arXiv 2608.24069 (25 Aug) is the first systematic study of adversarial signal propagation in multi-agent trading systems, and it deliberately restricts the attacker to what is practically reachable, the source data and prompts, rather than assuming privileged access to internals. It decomposes a widely used pipeline into Analyst, Researcher, Trader and Risk Manager roles, pairs each with an interface-matched attack, and evaluates four communication topologies under data-level and agent-level attacks across five assets, two backbones and two target directions. The Adversarial Signal Preservation Score is the transferable idea: a per-topology measure of how far a corrupted signal survives toward the decision.
Source
↳ Follow the thread